Pdiscoformer has been accepted at ECCV 2024 as an oral presentation!
PDiscoFormer has been accepted at the main conference of ECCV 2024 as an oral presentation!
Happy to share that my new paper, titled 𝐏𝐃𝐢𝐬𝐜𝐨𝐅𝐨𝐫𝐦𝐞𝐫, co-authored with Diego Marcos, Cassio F. Dantas, and Dino Ienco, has been accepted at the main conference of the 𝐄𝐮𝐫𝐨𝐩𝐞𝐚𝐧 𝐂𝐨𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐨𝐧 𝐂𝐨𝐦𝐩𝐮𝐭𝐞𝐫 𝐕𝐢𝐬𝐢𝐨𝐧 (𝐄𝐂𝐂𝐕) 2024, scheduled to be held in Milano this October!
In our work, we explore computer vision methods that perform unsupervised part discovery. We introduce a novel method and training objective for this task using self-supervised vision transformers, achieving state-of-the-art results. Our model learns to discover consistent, discriminative parts that are useful for solving image classification tasks, taking a step towards inherently interpretable models.